Title: LERD: Latent Event-Relational Dynamics for Neurodegenerative Classification

Abstract:

Accurate diagnosis, screening, and monitoring of Alzheimer’s disease (AD) rely increasingly on the precise analysis of brain electrophysiology, as the condition significantly disrupts multichannel EEG dynamics. Despite this need, current methodologies often depend on opaque, black-box classifiers that fail to explicitly account for the latent event timing and inter-channel coordination underlying their predictions. To overcome these constraints, we introduce LERD, an end-to-end Bayesian latent event-relational dynamical system. This framework directly infers latent neural events and their relational structures from multichannel EEG data, operating without the need for event or interaction annotations. LERD integrates a continuous-time event inference module with a stochastic event-generation process to model flexible temporal patterns, while utilizing an electrophysiology-inspired dynamical prior to ensure principled learning. Our theoretical analysis establishes a tractable IVP-based KL regularizer and provides stability guarantees for the inferred relational dynamics. Comprehensive evaluations on synthetic benchmarks and two real-world AD EEG cohorts show that LERD consistently surpasses strong baselines. Furthermore, it produces physiology-aligned summaries of rates, timing, and graphs, facilitating the characterization of group-level dynamical differences.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC